target_feature requires embedded LLVM copy to be usable

[semarie]

It seems that target_feature requires the embedded LLVM copy to populate the available target_features.

target_features() in src/librustc_trans/llvm_util.rs iterates on whitelisted lists of features, and check for each if the feature is available, using llvm::LLVMRustHasFeature().

The function comes from src/rustllvm/PassWrapper.cpp, and will return always false if LLVM_RUSTLLVM isn't defined. So if the binded LLVM isn't the embedded copy.

I found it when building rustc 1.22 (prestable) as run-pass/sse2 failed. But firefox-57 depends on simd crate and simd uses target_feature: it makes firefox unbuildable in such situation.

As for distro, it is really a common thing to avoid embedded copy of code, it is really annoying.

[semarie]

cc @alexcrichton @nagisa

[semarie]

the error I have while building stock firefox-57 on openbsd with rustc 1.22:

  --> /data/semarie/repos/openbsd/ports/pobj/firefox-57.0/firefox-57.0/third_party/rust/simd/src/common.rs:17:10
   |
17 | use x86::sse2::common;
   |          ^^^^ Could not find `sse2` in `x86`

[ranma42]

The limitation was introduced in 881a724

EDIT: to fix compilation after #43492

[cuviper]

But firefox-57 depends on simd crate and simd uses target_feature:

Isn't simd a nightly-only crate? How can Firefox be using it? It's present in third_party/rust/simd, but I think that's only as a vendored optional dependency for crates like regex.
(It does not appear to be used in Fedora's build of Firefox 57, at least...)

Anyway, AFAIK MCSubtargetInfo::getFeatureTable() and getCPUTable() are currently the only API changes in rust-lang/llvm. This wasn't a big deal when they were only informational, used for rustc --print target-features --print target-cpus. But it's more painful if these now affect actual code use. Have these APIs been proposed upstream to LLVM?

[semarie]

it seems the issue was already present with rust-1.21.0: freebsd already saw it: freebsd bug 223300

at openbsd, rust-1.21.0 was compiled with embedded LLVM. this one with system LLVM. so I only saw it now.

about firefox, it seems that firefox explicitely enable the build with simd for performance reason. see firefox 1261841

[semarie]

firefox enable nightly features for simd: config/rules.mk#875

[cuviper]

Then Firefox is cheating -- but you still have the option to build it without simd, it seems.

[semarie]

with #49428, the problem should be solved. thanks.